Blogs

A random mini experiment applying logit-lens probing to InternVL3 on our aerial imagery dataset

I spent the pandemic building things, debugging everything myself and falling in love with programming. Years later, AI has made coding dramatically easier and I'm starting to notice that some of what made it satisfying has quietly disappeared.

Introducing subjective questions impacted participant performance, revealing the gap between traditional auto-graded metrics and actual problem-solving skills.